An adaptive method of emergency department syndromic surveillance to nowcast the frequency of presentations that will have a severe 28-day outcome following influenza or COVID-19 infection: a retrospective analytical record linkage study

Abstract

During epidemics, emergency department (ED) syndromic surveillance of patient arrivals provides timely but nonvirus-specific assessment of epidemic intensity. Surveillance of severe infection outcomes (intensive care admission or death) is less timely because outcomes can take weeks to occur. Time series models can be used to estimate the frequency of severe infection outcomes due to viruses. We developed and evaluated daily time series modeling applied to linked ED, infection, and outcomes data from Australia to better predict population and health system burden during acute respiratory virus epidemics. In retrospective daily surveillance emulation, generalized additive models nowcasted (short-term forecast) the frequency of ED arrivals attributable to each of influenza and COVID-19 that will have a severe infection outcome within 28 days. Daily nowcasts spanned days -29 to -4 from each date for which surveillance was emulated. To validate the method, nowcasts were compared with subsequently observed severe infection outcome frequencies for December 2021 through February 2023. During this period, the mean daily day-4 nowcast error was 2.7 (34.2%), compared with 3.5 (43.8%) if outcomes known at day-1 were used. With increasing real-world data availability, this method could improve rapid, automated epidemic assessment for timely public health action.

Introduction

Emergency department (ED) syndromic surveillance systems are used by health jurisdictions to detect and monitor public health hazards, outbreaks, and epidemics. ¹ Internationally, their value in providing rapid information on the timing and severity of acute respiratory infections, including influenza and COVID-19 due to the SARS-CoV-2 virus, is increasingly recognized.^1-3

Annual epidemics of seasonal influenza and more frequent waves of SARS-CoV-2 infections, sometimes concurrent, are now common, and subject to global surveillance.^4,5 The frequency with which people experience a severe infection outcome is important for assessment of epidemic risk, population impact and for planning health and other service provision.

The World Health Organization (WHO) previously developed guidance for pandemic influenza severity assessment (PISA), with the intention it be applied to both pandemic and seasonal influenza epidemics. PISA has four severity dimensions: “transmissibility” of the virus, “seriousness of disease,” “morbidity and mortality,” and “impact on health care capacity.” These dimensions also apply to assessing the impact of the COVID-19 pandemic.^6,7

Time series models are often used to retrospectively obtain single time period estimates of morbidity and mortality from influenza or other epidemic respiratory infections because hospital diagnosis or cause of death-based outcome data may inaccurately estimate the incidence of severe infection outcomes, and testing of patients for the infection is usually incomplete.^8,9 Commonly, the models incorporate incidence of virologically detected infections as predictor variables and incidence of death or hospitalization as an outcome.^9-12 Simpler methods based on Serfling’s original trigonometric time series regression approach¹³ without virus-specific predictor variables, continue to be used for routine, continuous surveillance of excess epidemic mortality in systems like EuroMOMO for 25 countries.¹⁴ The Serfling approach, however, cannot discriminate outcomes attributable to individual viruses.

New South Wales (NSW), Australia, with a 2024 population of 8.52 million, operates near real-time ED syndromic surveillance¹⁵ that supports routine COVID-19 and influenza epidemic intelligence.¹⁶ The ED surveillance provides timely assessment of hospital demand during an epidemic but outcomes are only known as at discharge from the ED, and infection test result records are not available in ED records.¹ Thus, a particular challenge is distinguishing which infection is driving an increase in presentations.¹⁷

Another challenge in ED surveillance is that severe infection outcomes often occur some days after arrival at hospital, which delays knowledge of the frequency of those outcomes and of PISA-based estimation of seriousness, morbidity and mortality, and health care impact assessments.¹⁸ The severity of infection outcomes can also vary over time, depending on the prevalent virus variants, or due to epidemic control measures.^19,20

This retrospective analytical study used a novel database to demonstrate application of three methodological advances to daily ED syndromic surveillance of acute-respiratory infection-like (“ARI-like”) ED presentations. The aim was to improve timeliness, virus-specificity, and seriousness measures during acute respiratory infection epidemics, with influenza and COVID-19 as examples. First, health system dataset records were linked to provide outcomes following ED presentation. Second, daily time series models were automated to estimate the frequency of severe virus-attributable patient infection outcomes. Third, the models facilitated nowcasting (short-term forecasting) of recent severe outcome frequencies for each virus prior to those outcomes occurring. The seriousness outcome of interest is intensive care unit (ICU) admission or death within 28 days of presentation.

Methods

Data source

The anonymized Pandemic and Epidemic Assessment of Risk using Linked (PEARL data) database was used as the basis for this analysis. The PEARL database includes ED presentation records from all 184 public hospitals that report to the NSW ED Data Collection (EDDC).²¹

NSW ED surveillance uses syndromes formed from groupings of related principal ED diagnoses assigned at discharge from the ED.¹⁵ The diagnoses can represent symptoms or specific conditions. The PEARL database includes only ARI-like syndromes: acute respiratory infection, fever, unspecified infection, breathing problems (excluding asthma), or sepsis. These categories reflect the spectrum of diagnoses in the state’s electronic medical record systems that may be used by ED physicians for patients presenting with an acute respiratory infection.^22,23 Some of these diagnoses could be due to a nonrespiratory infection, but no other information is available to further refine the group. The included diagnoses were previously published.¹⁹

The PEARL database additionally includes person-based, probabilistic record linkage²⁴ to confirmed influenza and COVID-19 infections routinely reported to the statewide Notifiable Conditions Information Management System (NCIMS) by all public and private microbiology facilities.²⁵ In NCIMS, influenza notifications require laboratory definitive evidence (confirmed), while COVID-19 notifications require laboratory definitive evidence (confirmed) or a rapid antigen test (probable) result.²⁵ Linkage is also with inpatient hospitalization records, including ICU admission status, from the NSW Admitted Patient Data Collection and death records from the NSW Registry of Births, Deaths, and Marriages,²¹ with inclusion limited to admission episodes commencing, or deaths occurring, within 28 days following the ED arrival date.

The ED surveillance data stream considered in this study is daily frequencies of ARI-like presentations and the outcome of interest is a 28-day outcome of ICU admission or death following initial ED presentation, called “severe ARI-like presentations” here. ICU admission was identified from linked inpatient episodes with > 0 hours recorded in an ICU. A death registration can include death in hospital or the community.

Confirmed influenza or COVID-19 infection was based on presence of a linked infection notification with a specimen collection date in the interval -28 to +3 days from initial ED arrival. This interval allowed sufficient time for an infection to develop symptoms or complications leading to ED presentation, and up to 3 days following arrival at hospital for testing to occur while limiting the risk of including hospital-acquired infections.

The record linkage permitted repeat ED presentations with ARI-like syndromes for the same patient to be recognized. Any repeat presentations within a 28-day period were ignored, to avoid multiple counting of the same patient. The arrival date of the initial presentation determined the beginning of the 28-day follow-up period, with day 1 of follow-up being the day after presentation.

Analysis

In real world daily syndromic surveillance operation, a health department would analyze the latest available ED data to understand current disease trends. We call the day on which surveillance analysis is conducted the “surveillance date.” Each new day brings a new surveillance date. For a given surveillance date, the analysis would typically use baseline ED data available up to and including the previous day (which we call “day-1” here). The baseline period provides expected values against which recent trends are referenced. With this in mind, we present a procedure where baseline data are used to nowcast recent daily counts of severe ARI-like presentations attributable to influenza or COVID-19, that is, the daily frequency of those infections with a severe infection outcome.

For each surveillance date, the analysis proceeded as summarized in Figure 1. First, the baseline and nowcast periods were determined and data for those were extracted from time series generated from the PEARL database. The baseline period was set at two years to provide sufficient virus activity for model estimation during periods when incidence was low, while avoiding too much influence from artifacts of patient testing frequency, which have changed over time in Australia.²⁶ The baseline also needed to include only ED presentations with known 28-day outcomes, so the most recent 28 days prior to the surveillance date were excluded. Thus, the baseline was days -758 to -29 relative to the surveillance date (Figure 1).

Figure 1.

Analysis steps for a single surveillance date and schematic representation of the baseline and nowcast periods relative to the surveillance date (not to scale).

The necessary exclusion of the previous 28 days determined the nowcast period, which commenced at day -28. Because an inclusion criterion for the specimen date of a linked virus detection was up to three days following ED presentation, the nowcasts would use incomplete data if nowcasting days -3 to -1. Therefore, the nowcast period was fixed at days -28 to -4 (Figure 1).

Second, a Gaussian generalized additive model (GAM)²⁷ was fit to the baseline time series of daily counts, by ED arrival date, of severe ARI-like presentations as the outcome variable. The model equation is:

(1)

where the outcome is the daily count of ARI-like presentations to the ED on a given ED arrival date (). The parametric component of the GAM includes weekday and and , which are the counts of ARI-like presentations arriving by ambulance on that date that have a notified influenza or COVID-19 infection. The spline term of ED arrival date accounts for variation in the unobserved background incidence of severe ARI-like presentations attributable to causes other than those represented in the parametric terms. The model error term is .

The daily frequency of patients arriving by ambulance was used as the predictor variable because mode of arrival can be known on the same day as ED arrival, is less likely to be influenced by changes in availability of other health services that would affect patients with milder illness, and ambulance arrival was associated with severe infection outcomes for both viruses in a recent study.¹⁹ For COVID-19, counts were set to zero for all days prior to the first occurrence of a linked case in 2020.

Rather than use automated smoothing, the spline of ED arrival date had knots fixed at the beginning and end of the series and at the central date of each season of the year, that is, mid-January, mid-April, mid-July, and mid-October. Season-based knots were used because background incidence usually varies seasonally in NSW.²³ Knots were fixed to ensure the same knots were used for successive model iterations, ensuring consistency of estimation in successive surveillance days. The inclusion of a flexible spline rather than cyclic harmonic terms that are often used in similar analyses, allowed more flexibility to handle disruption in seasonality, such as occurred for many respiratory infections during the COVID-19 pandemic in 2020 and 2021 in NSW.²⁸

Finally, for nowcasting, model coefficients were used to estimate severe ARI-like presentations attributable to each of influenza and COVID-19. For a given ED arrival date, the coefficient for each virus-linked ED ambulance arrival variable was multiplied by the respective observed ARI-like ambulance arrival counts for that date. The 95% confidence limits of estimates and nowcasts were obtained in the same way except the coefficients were replaced by their 95% confidence limits. Values were rounded to the nearest integer. For visualization, the estimates were calculated and presented for both the baseline and nowcast periods.

Analysis was conducted in SAS Enterprise Guide version 8.3 with Base SAS version 9.4. The GAM, implemented with the SAS GAMPL procedure, assumed normally distributed errors. GAMPL implements thin plate regression splines²⁹ for the nonparametric smoother and uses penalized likelihood estimation. Model assumptions were checked using plots of residuals against predicted values, autocorrelation functions (SAS ARIMA procedure) and QQ-plots (Univariate procedure). Model smoothing parameters were selected using METHOD = PERFORMANCE in the GAMPL model statement, and the optimization technique selected was TECHNIQUE = NMSIMP, which provides Nelder–Mead simplex optimization. If convergence failed, the algorithm was designed to automatically repeat the model using Newton–Raphson optimization (TECHNIQUE = NEWRAP).

Emulation and evaluation

To emulate real-world operation, surveillance analysis was retrospectively replicated and nowcasts produced over 2250 surveillance dates from January 1, 2017 through February 28, 2023.

To examine the performance of the nowcasts, we used the SARS-CoV-2 Omicron variant epidemic period (December 1, 2021 to February 28, 2023). The period was selected because universal testing of patients for COVID-19 infection in the state’s hospitals during the pandemic meant that ascertainment of infection was high and by then social lockdown and other strict measures had been eased,³⁰ giving a more consistent susceptibility and epidemiological profile, over time, of infected persons. We were unable to evaluate the performance for seasonal influenza because there is no universal testing policy for influenza infections in NSW so comparative reference counts of confirmed infections were unavailable.

Performance was assessed using the mean of the daily absolute difference, the root mean square error (RMSE) and coverage probability of the 95% confidence intervals for each day -4 nowcast and the subsequently observed count of severe ARI-like presentations, with 28 days’ follow-up. These were compared with the difference that would be observed if the outcomes known at the surveillance date (with the available 3 days’ follow-up) were used as the 28-day follow-up estimate instead.

Sensitivity analyses

To assess the effect of increasing or decreasing the baseline model-fitting period, the procedure was repeated with a 1- and 3-year baseline. To assess a simpler choice of independent variables, the analysis with a 2-year baseline was repeated using all ARI presentations with a linked virus notification instead of those arriving by ambulance.

Results

Descriptive statistics of the data used

During December 5, 2014 through February 28, 2023, there were 86 960 severe ARI-like presentations to 176 included facilities. Among these, patients aged 65-84 years accounted for the largest proportion (43.9%) of the presentations, followed by patients aged 84 years or more (22.6%), while under-5-year-olds accounted for 6.0%. Females comprised 43.7% of patients. Seventy per cent of presentations arrived by ambulance, 9.8% were triaged as immediately life threatening, and 48.4% were triaged as imminently life threatening. Sepsis (26.4%), breathing or respiratory problems excluding asthma (25.5%), and pneumonia (22.3%) were the most common syndrome categories (Appendix S1, Table S1).

During the same period (December 5, 2014 to February 28, 2023), the mean daily count of the outcome variable, severe ARI-like presentations, was 28.9 (median = 28). There were 15 918 ARI-like presentations arriving by ambulance with a linked influenza infection record, giving a mean daily count for the influenza predictor variable of 5.3 (median = 1). For the equivalent COVID-19 predictor variable (from January 1, 2020), the mean was 30.8 (median = 10) (Table 1).

Table 1.

Summary statistics for the daily ED presentation frequency variables used in the surveillance models, December 5, 2014 through February 28, 2023.

Variable (all daily frequencies)	Mean	Median	Q1	Q3	Minimum	Maximum
Model outcome: severe ARI-like	28.9	28	23	33	11	74
Predictor: any ARI-like arriving by ambulance and confirmed influenza	5.3	1	0	4	0	89
Predictor: any ARI-like arriving by ambulance and confirmed COVID-19a	30.8	10	0	51	0	300

^aEstimated from January 1, 2020 only. Q: quartile.

Surveillance emulation

Figures 2 and 3 show example emulations for surveillance dates of August 10, 2017 and August 1, 2022, respectively. The 2017 example (Figure 2) was early in the severe annual influenza season for that year and during the pre-COVID-19 period. The nowcast period highlights the rapid increase in severe ARI-like presentations, exceeding levels of prior years. The 2022 example (Figure 3) occurred during the COVID-19 pandemic when seasonal influenza re-appeared after transmission ceased in Australia during the first two years of the pandemic.²⁶ The model was able to discriminate presentations attributable to each virus. Regarding model assumptions for these two examples, the residual versus predicted, autocorrelation, and QQ-plots of the model residuals supported the choice of a Gaussian model and the independence of the residuals (Appendix S1, Figures S1 and S2).

Figure 2.

Example surveillance analysis example during a severe influenza season prior to the emergence of COVID-19. Emulated surveillance date is August 10, 2017 with baseline (model fitting) period July 14, 2015 through July 12, 2017. Estimates and nowcasts include 95% confidence interval bands (shaded). As this period was prior to the COVID-19 pandemic, and COVID-19-linked presentation counts were zero, the model did not estimate any COVID-19 outcomes. Abbreviations: ED: emergency department; ARI-like: acute respiratory infection-like.

Figure 3.

Example surveillance analysis during the COVID-19 pandemic coinciding with the recurrence of seasonal influenza for the first time since the beginning of the pandemic. Emulated surveillance date is August 1, 2022 with baseline (model fitting) period July 4, 2020 through July 3, 2022. Estimates and nowcasts include 95% confidence interval bands (shaded). Abbreviations: ED: emergency department; ARI-like: acute respiratory infection like.

An important consideration for daily, automated surveillance analysis is the ability of the model fitting to successfully complete, that is, converge on a result, thereby avoiding gaps in surveillance. All 2250 GAM models achieved convergence, that is, were able to produce estimates, using Nelder–Mead simplex optimization in all but three (0.13%) of 2250 emulated surveillance days. In all three convergence failures, automatically re-running the model with Newton–Raphson optimization achieved convergence.

Another important consideration is the precision of the estimates and nowcasts produced by the models. For influenza, based on a two-sided P < 0.05, the model coefficients were precise enough to produce statistically significant nowcasts during the 2017 through 2019 influenza seasons. For seasonal influenza that returned from the end of March 2022 after transmission ceased in the first two years of the COVID-19 pandemic, statistically significant nowcasts were consistently available within 6 weeks, by May 10, 2022 (Appendix S1, Figure S3).

For COVID-19, the first nowcasts were produced for surveillance dates from March 17, 2020, coinciding with the commencement of the initial, brief outbreak that ended mid-April 2020 (Figure 4). Coefficients producing the nowcasts were statistically significant for surveillance dates from April 17 through May 2, 2020, following the initial outbreak (Appendix S1, Figure S3). Transmission remained mostly suppressed until a large outbreak due to the SARS-CoV-2 Delta variant from June 16 through November, 2021³¹ (Figure 4). Coefficients continued to be significant from then until the end of the study period, and decreased around July 31, 2021, 6 weeks from the start of the Delta outbreak. Coefficients then remained stable from around September 24, 2021 until the end of the study period (Appendix S1, Figure S3) including through the Omicron variant period that started late November 2021.

Figure 4.

Day-4 nowcast counts compared with subsequently observed daily counts of severe acute respiratory infection-like ED presentations with a linked influenza or COVID-19 infection, for each daily surveillance iteration, by year. Notes: Prior to 2020, COVID-19-linked presentation counts were zero. Therefore, the model did not estimate any COVID-19 outcomes during that period. Abbreviation: ED: emergency department.

Figure 4 compares, for each surveillance iteration, the day-4 nowcast with the subsequently observed daily counts of severe ARI-like presentations that had a linked infection record, for the same ED arrival date. In 2017, the nowcast influenza level was higher than the observed level. In 2022, during the period in which seasonal influenza returned after the first two years of the COVID-19 pandemic, nowcasts were substantially higher than observed counts with the model baselines including little influenza activity and confidence intervals were wide (Figure 3). Again, for COVID-19, nowcasts were substantially higher than observed values during the relatively brief COVID-19 outbreaks in the first two years of the pandemic. From April 2022, the COVID-19 nowcasts visually remained close to the observed counts (Figure 4).

The performance assessment compared day-4 nowcasts with subsequently observed COVID-19-attributable severe ARI-like presentations with 28 days’ follow-up during the 451 days of the included Omicron variant epidemic period. During the period, the mean daily observed count was 7.9 and the mean daily absolute difference was 2.7, or 34.2% of the mean observed daily count. The RMSE was 4.08, or 51.6% of the mean observed counts and the coverage probability of the 95% confidence intervals was 30.2%. Using, instead, as the estimate the observed count available at the surveillance date, when only three follow-up days were available, the mean daily absolute difference was 3.5, or 43.8% of the mean observed daily count with 28 days’ follow-up [RMSE: 4.96 (62.4%)].

Sensitivity analyses with 1- and 3-year baselines also gave mean absolute differences of 2.7, and RMSE of 4.07 (51.5% of the mean observed counts) and 3.91 (49.5%), respectively. The respective coverage probabilities were 33.9% and 31.9%. The 1-year baseline produced more unstable estimates for both influenza (especially 2019 and 2022) and COVID-19 following periods of sparse virus activity (Appendix S1, Figure S4). The 3-year baseline produced consistently stable estimates throughout the analysis period (Appendix S1, Figure S5). The sensitivity analysis using any ARI-like presentation linked to a virus notification as the independent variable in the model (not just those arriving by ambulance) provided poor model fit (see Appendix S1 for further details).

Discussion

Using linked data sources and methods typically applied in one-off estimation of influenza and other virus-attributable infection outcomes in populations, we demonstrated an adaptive surveillance method to provide timely nowcasts of the frequency of virus-specific emergency department presentations that will have a severe infection outcome of intensive care admission or death within the 28 days following arrival at the ED. Nowcasting was facilitated using infection notifications linked to the most recent ARI-like ED presentations that arrived by ambulance, indicating presentations with the potential for a severe outcome.

Most studies estimating population-level virus-attributable outcomes lack a reference standard for comparison, and estimates can vary widely.³² We were fortunate to have complete testing of hospitalized patients in our state during the COVID-19 pandemic and the Omicron period had continuous COVID-19 transmission that allowed us to assess the accuracy of our approach. During that period, the mean of the absolute nowcasts differences for the day-4 nowcasts was 34% of the mean subsequently observed values with 28 days’ follow-up. This was lower than the 44% using observed follow-up data known at the surveillance date. From April 2022, when the model would have adapted to the more consistent population susceptibility and epidemiology of the Omicron variant epidemic starting in December 2021, the day-4 nowcasts were, visually, much closer to the observed values. Low coverage probabilities, around 30%, for the nowcasts’ confidence intervals highlight the need for methods to be developed for prediction interval estimation to take into account the variation in observed values rather than their mean.

The marked reduction in the COVID-19 coefficient from late July 2021, early in the Delta variant outbreak, suggests a reduction in the frequency of severe outcomes among patients with COVID-19 presenting to ED by ambulance. This is likely explained by the rollout of the COVID-19 vaccine from February 2021, which included early prioritization of older and other especially vulnerable persons, such as aged care residents, and rapid increases in overall population coverage during the Delta outbreak.^30,33 Changing severity of SARS-CoV-2 variants, as observed in another NSW study,²⁰ may have also contributed.

Sensitivity analyses (see Appendix S1) showed that a 3-year baseline will provide a similar accuracy compared with a 2-year baseline and a higher proportion of statistically significant nowcasts than shorter baselines. It also produced stable and statistically significant estimates in 2022 following the long period without influenza transmission during the first two years of the COVID-19 pandemic. However, the 3-year baseline requires more historic data and could be more sensitive to changes in patient testing frequency. The 1-year baseline was too short to produce consistent and statistically significant influenza nowcasts for large parts of the study period.

As this was a retrospective study, we were able to include COVID-19 in the surveillance algorithm for surveillance occurring before its emergence in 2020. This parallels the inclusion of an unknown “Disease X.” Disease X is a placeholder for a future, as yet unknown, pandemic pathogen that is recognized in current pandemic preparedness planning.³⁴ If, as happened with SARS-CoV-2,³⁵ diagnostic testing for Disease X is rapidly developed, then this surveillance approach could be feasible within months of a new virus’ emergence.

The applicability of the approach in other settings depends on availability of suitable linked medical records from which ARI-like syndromes could be extracted, and recording of virus detections. The WHO continues to facilitate introduction of national standardized respiratory virus testing centers assisting uniform virus detection and reporting.³⁶ Record linkage would be unnecessary if, as recommended by the WHO, digital health-care records were comprehensive, routine, standardized, and interoperable, and the number of countries with this capability is growing.^37-39

Limitations include the unavailability of data for hospitals serving one local government area in Sydney and the few private EDs in NSW, combined representing <3% of NSW EDs. The approach should be considered a proof of concept, since the methods assumes laboratory detections and severe outcomes are almost immediately available for surveillance. Currently, NSW public hospitals have electronic medical records from which real-time patient status, including death, could be collected, but only ED data is used in this way for routine ED surveillance.⁴⁰ Approximately one fifth of infection deaths occur outside hospitals,¹⁹ but in NSW, > 80% of deaths were registered within 28 days, even in pandemic years.⁴¹ Given the baseline excludes the latest 28 days, the small loss of data at the end of the baseline period would only lead to a slight bias. Intensive care unit capacity-related admission policies could lead to biased estimation of severe outcomes.

In NSW, there was strong testing capacity and universal testing of hospital patients throughout the COVID-19 pandemic, with state pathology services processing > 6 million tests during 2020 through 2021.⁴² Nevertheless, testing capacity was exceeded in early 2022 when the Omicron variant emerged.⁴³ Rapid COVID-19 antigen tests were reportable from January through October 2022 and were included in notified infections, but may have been under-reported. Inclusion of incidental infections in patients presenting with, rather than for, COVID-19,⁴⁴ would be mitigated by our restriction to ARI-like ED diagnoses.

For simplicity, we chose a Gaussian model. While our data met the modeling assumptions for the Gaussian approach, this may not be the case in other settings. Further development will be required to extend the approach to log-linear negative binomial or quasi-Poisson models used for count data. Simulations for testing operating characteristics in alternative and diverse settings could be a subject for future research, as would forecasting the frequency of severe infection outcomes over a longer time frame.

In conclusion, we demonstrated a daily surveillance methodology to nowcast the recent frequency of ED presentations attributable to specific acute respiratory infections that will have a severe outcome. While the method cannot replace intense, active surveillance needed in the first months of community transmission of a new or re-emerging infection, it shows promise for improving rapid severity assessment for epidemic intelligence once community circulation is established. As electronic healthcare records become more mainstream and interconnected, richer surveillance opportunities like this should eventuate.

Supplementary material

Supplementary material is available at the American Journal of Epidemiology online.

Conflicts of interest

None declared.

Funding

David Muscatello and Alexandra Hogan are supported by an Australian National Health and Medical Research Council (NHMRC) Investigator Grant (DJM: APP1194109; ABH: APP2009278). The contents of the published material are solely the responsibility of the administering institution, a participating institution, or individual authors and do not reflect the views of the NHMRC.

Data availability

Restrictions apply to the availability of these data. The data that support the findings of this study can be requested from The Centre for Health Record Linkage (CHeReL: https://www.cherel.org.au/), with permission for release required from the New South Wales Ministry of Health.